- Apr 20, 2009
- 694
- 697
- Awards
- 8
- First Name
- Ryan
TLDR: “Data matters more than the algorithm." - Professor Iyad Rahwan, MIT
Meet Norman.
Norman is the stuff bad sci-fi movies are made of, and the real result of programming a machine to learn from a fixed set of massively skewed data.
Rather than give Norman, the name of a recent AI project at MIT, a well-rounded set of images to train his database of virtual experience, the researchers at MIT gave him the deepest, darkest, and most deplorable subs of Reddit. Then they asked him to tell them what he saw in several Rohrshach inkblots. The result is as predictable as it is disturbing.
Take this image below:
Where conventional, well-rounded, AI models return "a close up of a vase with flowers" to describe this image, Norman returns "a man is shot dead."
...and this one?
Standard AI returns "a black and white photo of a small bird." What does Norman see? "Man gets pulled into dough machine.”
If you'd like to read more of the ramblings of a psychopathic AI, here is the complete article from MIT on Norman: http://norman-ai.mit.edu/ It's a quick read.
"...when people talk about AI algorithms being biased and unfair, the culprit is often not the algorithm itself, but the biased data that was fed to it."
While there are many interesting anecdotal applications and personal references of this AI experiment, I think the two most intriguing for the automotive industry may be these:
1. How do we limit our own confirmation bias?
2. Is the data skewed?
Norman serves as a reminder to question the personal confirmation bias that results from "knowing what has worked well in the past" at the exclusion of what may be required to adopt and adapt in the future. It is human instinct to view our current situation and decisions through the lens of our own past experience; that is a trait learned by machines as well.
The quote from the TLDR above is also instructive. It isn't about the strength of the algorithm, it is about the strength of the data. The best "shiny object" algorithms will return morbid success rates without well-rounded and complete data.
What do you think Refresh? How do you plan to keep your buyers, marketing managers, desk, and dare I say it, General Managers and Dealers, from being the Norman of automotive retail by relying on too little data and too much personal experience?
Meet Norman.
Norman is the stuff bad sci-fi movies are made of, and the real result of programming a machine to learn from a fixed set of massively skewed data.
Rather than give Norman, the name of a recent AI project at MIT, a well-rounded set of images to train his database of virtual experience, the researchers at MIT gave him the deepest, darkest, and most deplorable subs of Reddit. Then they asked him to tell them what he saw in several Rohrshach inkblots. The result is as predictable as it is disturbing.
Take this image below:
Where conventional, well-rounded, AI models return "a close up of a vase with flowers" to describe this image, Norman returns "a man is shot dead."
...and this one?
Standard AI returns "a black and white photo of a small bird." What does Norman see? "Man gets pulled into dough machine.”
If you'd like to read more of the ramblings of a psychopathic AI, here is the complete article from MIT on Norman: http://norman-ai.mit.edu/ It's a quick read.
"...when people talk about AI algorithms being biased and unfair, the culprit is often not the algorithm itself, but the biased data that was fed to it."
While there are many interesting anecdotal applications and personal references of this AI experiment, I think the two most intriguing for the automotive industry may be these:
1. How do we limit our own confirmation bias?
2. Is the data skewed?
Norman serves as a reminder to question the personal confirmation bias that results from "knowing what has worked well in the past" at the exclusion of what may be required to adopt and adapt in the future. It is human instinct to view our current situation and decisions through the lens of our own past experience; that is a trait learned by machines as well.
The quote from the TLDR above is also instructive. It isn't about the strength of the algorithm, it is about the strength of the data. The best "shiny object" algorithms will return morbid success rates without well-rounded and complete data.
What do you think Refresh? How do you plan to keep your buyers, marketing managers, desk, and dare I say it, General Managers and Dealers, from being the Norman of automotive retail by relying on too little data and too much personal experience?
